Method
When two servitors seem plausible, compare the problem mechanism, what evidence each model predicts and which wrong selection would be easiest to detect.
Practice rule
Every interpretation should preserve function, state uncertainty and produce either an observable check, a bounded action or a documented reason to stop.
Audit rule
Record what was known before the operation, what was inferred, what happened and what did not happen. Do not rewrite the prediction after the event.